MLS-C01 Hard Questions — 25 Difficult Questions

Question 1

An ML team is deploying a model to a SageMaker endpoint for real-time inference. The model is large (2 GB) and requires GPU for low-latency inference. The team wants to minimize cost while maintaining a response time of under 200 ms. Which instance configuration and SageMaker feature would be best?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Use a GPU instance (ml.p3.2xlarge) with SageMaker Elastic Inference.. The best choice. By using a GPU instance (ml.p3.2xlarge) combined with SageMaker Elastic Inference, the team can attach a fractional GPU accelerator to the instance, providing the necessary GPU compute for low-latency inference while minimizing cost compared to using a full GPU instance alone. Option B (batch transform) is designed for offline inference on large datasets, not real-time. Option C (serverless inference) typically uses CPU instances and may incur cold starts, which can exceed the 200 ms response time requirement. Option D (multi-model endpoint) is optimized for hosting multiple models on a single instance, but does not directly address GPU acceleration or cost minimization for a single large model.

Answer

Use a batch transform job on GPU instances.

Answer

Use a serverless inference endpoint with a CPU instance.

Answer

Use a multi-model endpoint on a GPU instance.

Question 2

A data scientist is training a neural network on Amazon SageMaker. The network has many layers and the training is very slow. The scientist suspects that the gradients are vanishing. Which technique is most specifically designed to mitigate the vanishing gradient problem?

Accepted Answer

Use batch normalization.. Batch normalization helps by normalizing the activations, which reduces the problem of vanishing/exploding gradients. Dropout is for regularization. Data augmentation increases data. Gradient clipping deals with exploding gradients, not vanishing.

Answer

Use gradient clipping.

Answer

Use data augmentation.

Answer

Use dropout layers.

Question 3

A financial services company needs to build a data lake on Amazon S3 that meets regulatory requirements for data retention and encryption. Data must be encrypted at rest and in transit, and access must be audited. The data lake will be queried by Amazon Athena and Amazon Redshift Spectrum. Which combination of actions should be taken?

Accepted Answer

Enable S3 default encryption with SSE-KMS and enable AWS CloudTrail for S3 data events.. Regulatory requirements for encryption at rest and audited access are best met by enabling S3 default encryption with SSE-KMS, which provides customer-managed key control and auditability via CloudTrail, and enabling CloudTrail data events for S3, which logs object-level access. This combination satisfies encryption and auditing mandates for Athena and Redshift Spectrum queries.

Answer

Use IAM policies to control access and enable S3 server access logging.

Answer

Use SSL/TLS for all connections and enable S3 versioning.

Answer

Enable S3 default encryption with SSE-S3 and use S3 access logs.

Question 4

A data scientist is training a model using Amazon SageMaker. The training dataset is 500 GB and is stored in S3. The data scientist wants to use Pipe input mode to stream data directly from S3 to the training container. However, the training job fails with an error indicating that the container cannot read the data. What is the most likely cause?

Accepted Answer

The training algorithm does not support Pipe mode. Pipe mode in SageMaker streams data directly from S3 to the training container via a FIFO pipe, but the training algorithm must be specifically designed to read from a pipe (e.g., via stdin or a named pipe). If the algorithm expects random access to files or uses libraries that require seekable file handles (like many image-processing or custom Python scripts using `open()`), it will fail because a pipe is a sequential, non-seekable stream. SageMaker’s built-in algorithms like BlazingText and XGBoost support Pipe mode, but custom algorithms often do not unless explicitly coded for it.

Answer

The training instance does not have enough memory

Answer

The IAM role does not have s3:GetObject permission

Answer

The data is compressed and Pipe mode cannot handle compressed data

Answer

The S3 bucket is in a different Region

Question 5

A machine learning engineer is deploying a model for real-time inference using Amazon SageMaker. The model is a large ensemble that requires 8 GB of memory and 4 vCPUs. The expected traffic is 100 requests per second with a 200 ms latency requirement. Which instance configuration should they choose?

Accepted Answer

ml.c5.2xlarge (8 vCPU, 16 GB). (ml.c5.2xlarge) is correct because it offers 8 vCPUs and 16 GB memory, meeting both the compute and memory requirements while being cost-effective for CPU-based inference. Option A (ml.t2.medium) is wrong because it has only 2 vCPUs and 4 GB, insufficient for the 8 GB memory need. Option C (ml.p3.2xlarge) is wrong because it includes a GPU, which is unnecessary and overkill for this CPU-bound workload, and costs more. Option D (ml.m5.large) is wrong because it has only 2 vCPUs and 8 GB memory, lacking the required vCPUs and barely meeting memory.

Answer

ml.t2.medium (2 vCPU, 4 GB)

Answer

ml.p3.2xlarge (8 vCPU, 61 GB GPU)

Answer

ml.m5.large (2 vCPU, 8 GB)

Question 6

A data scientist is analyzing a dataset with a continuous target variable and suspects that the relationship between a predictor and the target is non-linear. Which THREE techniques can the scientist use to explore and model this non-linearity?

Accepted Answer

Add polynomial features (e.g., x^2, x^3) and check if model performance improves.. Options C, D, and E are correct. Adding polynomial features (e.g., x^2, x^3) allows a linear model to capture non-linear relationships. Decision tree regressors naturally model non-linear interactions between predictors and the target. A scatter plot with a LOESS smooth curve visually reveals non-linear patterns in the data. Option A (logistic regression) is incorrect because it is for binary classification, not for exploring non-linearity with a continuous target. Option B (Pearson correlation) only measures linear relationships, so it is not suitable for detecting non-linearity.

Answer

Apply logistic regression to binarize the target.

Answer

Compute the Pearson correlation coefficient between the predictor and target.

Question 7

A company is building a near-real-time dashboard using data from multiple sources. They need to aggregate millions of events per second with sub-second latency. The architecture must be fully managed and minimize operational overhead. Which service should they use for the aggregation layer?

Accepted Answer

Amazon Kinesis Data Analytics for Apache Flink.. Amazon Kinesis Data Analytics for Apache Flink is the correct choice because it provides a fully managed, serverless runtime for Apache Flink, which is designed for stateful stream processing at scale. It can aggregate millions of events per second with sub-second latency using exactly-once semantics and built-in checkpointing, meeting the near-real-time dashboard requirements without any infrastructure management.

Answer

AWS Lambda functions triggered by Kinesis Data Streams.

Answer

Amazon EMR with Spark Streaming.

Answer

Amazon Redshift with materialized views refreshed frequently.

Question 8

A data scientist is training a convolutional neural network (CNN) for image classification using Amazon SageMaker. The training loss decreases steadily but validation loss starts increasing after a few epochs. Which action should the data scientist take to address this issue?

Accepted Answer

Implement early stopping based on validation loss. The described behavior—training loss decreasing while validation loss increases—is a classic sign of overfitting. Early stopping monitors the validation loss and halts training when it stops improving (or starts to increase), preventing the model from memorizing noise in the training data. In SageMaker, this can be implemented using the `EarlyStopping` callback in the framework's estimator or by setting `use_early_stopping` to True in a built-in algorithm.

Answer

Increase the learning rate

Answer

Add more convolutional layers

Answer

Increase the batch size

Question 9

A company uses AWS Glue crawlers to populate the AWS Glue Data Catalog from Amazon S3. The data is partitioned by year/month/day/hour. The crawler runs every hour and adds new partitions. However, the data engineer notices that the crawler is taking longer to run as the number of partitions grows, and sometimes it misses new partitions. What is the most cost-effective and reliable way to address this?

Accepted Answer

Enable the crawler's partition index feature.. Enabling the crawler's partition index feature allows AWS Glue to quickly find new partitions without re-scanning the entire table, reducing runtime and improving reliability. Option B is incorrect because manually adding partitions is error-prone and does not scale. Option C is incorrect because the MSCK REPAIR TABLE command is a manual step that does not automate the process. Option D is incorrect because increasing the crawler's frequency does not address the underlying issue of scanning overhead and may lead to resource contention.

Answer

Manually add new partitions using ALTER TABLE ADD PARTITION in Athena.

Answer

Use the Athena MSCK REPAIR TABLE command after the crawler runs.

Answer

Increase the crawler's schedule to run every 30 minutes.

Question 10

A data scientist is tuning a gradient boosting model using Amazon SageMaker Automatic Model Tuning. The objective metric is AUC. The training job converges quickly but the final model has low AUC on the validation set. Which hyperparameter should the data scientist adjust to improve validation AUC?

Accepted Answer

Decrease the learning rate and increase the number of rounds. Decreasing the learning rate and increasing the number of rounds is the correct approach because a low learning rate forces the model to take smaller steps toward the optimum, reducing overfitting and allowing more trees to contribute to the ensemble. This combination often improves generalization and validation AUC when the training job converges too quickly, indicating that the model is overfitting or underfitting due to aggressive learning.

Answer

Increase the subsample ratio of training data

Answer

Increase the learning rate

Answer

Increase the maximum depth of trees

Question 11

A data scientist is training a neural network for image classification. The dataset has 50,000 images across 100 classes. The model uses a ResNet-50 architecture pre-trained on ImageNet. The training loss decreases rapidly, but validation loss starts to increase after 5 epochs. Which of the following is the most effective technique to address this?

Accepted Answer

Use data augmentation to increase the diversity of the training set. The rapid decrease in training loss followed by an increase in validation loss after only 5 epochs is a classic sign of overfitting. Data augmentation artificially expands the training set by applying random transformations (e.g., rotations, flips, crops) to existing images, which improves the model's generalization and reduces overfitting. This is the most effective technique among the options because it directly addresses the lack of diverse training examples without changing the model architecture or training hyperparameters in a way that could destabilize learning.

Answer

Increase the learning rate

Answer

Add more layers to the network

Answer

Use a smaller batch size

Question 12

Refer to the exhibit. A data scientist runs the above AWS CLI command to create a SageMaker training job using the built-in Linear Learner algorithm. The training job fails with an error. What is the most likely cause?

Accepted Answer

The S3 data type is AugmentedManifestFile, but Linear Learner requires RecordIO or CSV. The most likely cause is that the S3 data type is AugmentedManifestFile, but the Linear Learner algorithm requires RecordIO or CSV format. SageMaker built-in algorithms have specific input format requirements, and Linear Learner supports RecordIO-Protobuf or CSV, not AugmentedManifestFile. This mismatch causes the training job to fail.

Answer

The IAM role does not have sufficient permissions

Answer

The instance type ml.m5.large does not support the Linear Learner algorithm

Answer

The MaxRuntimeInSeconds is too short

Question 13

A financial services company is deploying a machine learning model for credit risk assessment. The model must have an inference latency under 200ms and must be able to handle up to 1000 transactions per second (TPS). The company wants to minimize costs. The model is a gradient boosting model implemented in XGBoost. Which SageMaker deployment option should the team choose?

Accepted Answer

Deploy the model on a SageMaker real-time endpoint with multiple instances behind a load balancer.. A SageMaker real-time endpoint with multiple instances behind a load balancer provides the sub-200ms inference latency and 1000 TPS throughput required for synchronous, low-latency predictions. XGBoost models are well-suited for real-time endpoints, and horizontal scaling with multiple instances ensures high availability and cost efficiency by matching capacity to demand.

Answer

Use SageMaker Batch Transform to process transactions in batches.

Answer

Use SageMaker asynchronous inference for queued requests.

Answer

Use SageMaker Serverless Inference for automatic scaling.

Question 14

A machine learning team is using Amazon SageMaker to train a model with a custom algorithm packaged in a Docker container. The training job fails with the error 'Error: Unable to locate sagemaker-training toolkit.' What is the MOST likely cause?

Accepted Answer

The container does not include the SageMaker Training Toolkit. The SageMaker Training Toolkit is a required library that provides the interface between SageMaker and custom Docker containers for training jobs. When a custom container does not include this toolkit, SageMaker cannot execute the training script or communicate with the container, resulting in the 'Unable to locate sagemaker-training toolkit' error. This is a mandatory dependency for any custom training container used with SageMaker.

Answer

The container does not have internet access to download dependencies

Answer

The instance type is incompatible with the container

Answer

The training role does not have permissions to access the container repository

Question 15

Which THREE AWS services can be used together to build a serverless data pipeline that ingests streaming data, transforms it, and loads it into Amazon Redshift for analysis?

Accepted Answer

Amazon Kinesis Data Firehose. Amazon Kinesis Data Firehose (C) is correct because it is a fully managed, serverless service that can ingest streaming data and deliver it directly into Amazon Redshift (via S3 staging and the Redshift COPY command), making it ideal for the load stage of the pipeline. Amazon Kinesis Data Streams (D) is correct because it provides the serverless, scalable ingestion layer for capturing and buffering streaming data before it is consumed or transformed downstream. AWS Lambda (E) is correct because it is a serverless compute service that can be triggered by Kinesis to transform records in flight before they are loaded into Redshift, satisfying the transform requirement without managing servers. Amazon EMR (A) is not correct here because it is a managed big-data cluster service (EC2-based) rather than a serverless component, and it is not needed for this streaming ingest-transform-load pattern. Amazon SQS (B) is not correct because it is a message queue for decoupling applications, not a streaming data ingestion or delivery service for Redshift pipelines.

Answer

Amazon EMR

Answer

Amazon SQS

Question 16

A company uses Amazon SageMaker to train a model using the built-in Linear Learner algorithm. The training data contains missing values in some features. What is the best practice for handling missing values with this algorithm?

Accepted Answer

Impute missing values using mean or median imputation. Mean or median imputation is the best practice because Linear Learner expects numeric input and cannot handle missing values natively; imputing with mean or median preserves the dataset size and avoids bias introduced by dropping rows. Median is preferred for skewed distributions, while mean works for roughly symmetric data. This approach is standard in SageMaker's own documentation and examples for Linear Learner.

Answer

Remove rows with missing values

Answer

Set missing values to zero

Answer

Use the `handle_missing` parameter in the algorithm

Question 17

A company uses Amazon Redshift for its data warehouse. The data engineering team needs to load 10 TB of data from Amazon S3 into Redshift every night. The team wants to minimize the load time and use the fewest number of COPY commands. The data is in CSV format and is partitioned by date in S3. Which approach should the team take?

Accepted Answer

Use a manifest file with a single COPY command.. Using a manifest file with a single COPY command is the most efficient approach because it allows Redshift to load data from multiple S3 objects (partitioned by date) in parallel, automatically splitting the workload across cluster nodes. This minimizes load time by leveraging Redshift's parallel processing without requiring multiple COPY commands or manual concatenation, and it avoids the overhead of additional services like AWS Glue for a straightforward bulk load.

Answer

Use multiple COPY commands, one per partition.

Answer

Concatenate all data into a single large file before loading.

Answer

Use AWS Glue to transform the data and then load into Redshift.

Question 18

A company is deploying a machine learning model using Amazon SageMaker. The model must be updated frequently without downtime. Which TWO strategies can achieve this? (Choose two.)

Accepted Answer

Use blue/green deployment with endpoint variants.. Amazon SageMaker supports blue/green deployment using endpoint variants, where you can deploy a new model version alongside the current one and then shift all traffic to the new variant once validated. This approach ensures zero downtime because the existing endpoint remains active during the transition, and traffic is switched atomically. Option E is correct because canary deployment with SageMaker allows you to gradually shift a small percentage of traffic to a new model variant, monitor its performance, and then ramp up to 100% if successful, all without interrupting the service.

Answer

Update the model artifact on the existing endpoint.

Answer

Delete the existing endpoint and create a new one.

Answer

Use rolling update with multiple instances.

Question 19

A bank is building a credit risk model using a large dataset with 500 features and 2 million samples. The dataset contains many categorical features with high cardinality (e.g., zip code, occupation). The model must be deployed on SageMaker and provide real-time predictions with low latency. They also need to explain individual predictions for regulatory compliance. Which approach is most appropriate?

Accepted Answer

Use a gradient boosting model with ordinal encoding for categorical features and use SageMaker's built-in XGBoost with SHAP. XGBoost with ordinal encoding and SHAP balances performance, latency, and explainability.

Answer

Use a linear model with target encoding for categorical features and deploy with SageMaker's built-in linear learner algorithm

Answer

Use a deep neural network with embedding layers for categorical features and use SageMaker's built-in Debugger for explanations

Answer

Use XGBoost with one-hot encoding for categorical features and deploy with SageMaker's built-in SHAP explainer

Question 20

A company is using SageMaker to train a large NLP model. The training job is taking too long due to high I/O wait time. The data is stored as CSV files in S3. Which optimization should the company implement to reduce I/O wait time?

Accepted Answer

Use SageMaker Pipe mode to stream data directly from S3. SageMaker Pipe mode streams data directly from S3 into the training algorithm without first writing it to the local disk, eliminating the I/O wait time caused by downloading and decompressing CSV files. This is the most effective optimization for high I/O wait during training because it bypasses the bottleneck of writing large datasets to the instance's local storage.

Answer

Convert CSV files to RecordIO format

Answer

Use SageMaker batch transform before training

Answer

Use SageMaker File mode with larger instance storage

Answer

Use SageMaker ShardedByS3Key data distribution